🦙 AB Simple Ollama Connection

Let's be honest about what this node is: it's a config router, not a worker. It doesn't talk to a model, it doesn't generate anything. What it does is save you from the exact failure mode that hits everyone running a few LLM nodes at once - five copies of the same Ollama URL scattered across a workflow, one of them wrong, and you burning an hour figuring out which.

If you're running a single 🦙 AB Simple Ollama node, skip this one entirely. The moment you have two or three - say, an Ollama node rewriting your prompt before a diffusion pass, plus one summarizing results afterwards - you wire this in and never type the URL again. It's the one place to change the server when you move Ollama to a new machine.

What it actually does

SimpleOllamaConnectionNode is deliberately thin. You feed it an ollama_url string, and it hands back the same URL, normalized. The run() method strips trailing slashes, rejects anything that isn't http:// or https://, insists on a host, and refuses URLs with query parameters or fragments. If you type garbage, the node's validation catches it before you even queue - the input socket glows red with the exact reason instead of failing mid-run.

Its one trick is the Connect Downstream button on the node's face. Click it and the frontend fetches your model list once from the server, then pushes it to every Simple Ollama node whose optional ollama_url_override input is wired to this node's ollama_url output. Downstream nodes keep their currently selected model if it still exists; if not, they fall back to the first available one and refresh their model info. Note that it overwrites each node's ollama_url widget with the shared URL - so this is where you centralize the connection, and the widget becomes a fallback.

The one wire that matters

The only input is ollama_url (the tooltip calls it a "Shared Ollama base URL for downstream Simple Ollama nodes"), and the only output is ollama_url. That's it. Connect the output to each Simple Ollama node's optional ollama_url_override input, and the shared URL wins whenever the override is connected.

Installing it

Via ComfyUI Manager, search comfyui-simple-ollama in the custom-node browser (it's on the Comfy Registry) and restart. Or manually:

cd ComfyUI/custom_nodes/
git clone https://github.com/a-und-b/ComfyUI_Simple_Ollama.git

Restart ComfyUI. No pip step - the pack leans on the Python stdlib plus PIL/numpy, which any ComfyUI already ships. Python 3.10+ is the floor. You do need Ollama itself running somewhere reachable: ollama serve on localhost, or any machine on your LAN that answers on port 11434.

Gotchas

• The default URL is a placeholder. The widget ships as http://192.168.1.x:11434 - literally with an .x in it. That string passes the node's validation (it's structurally a fine URL) but the Connect button will fail to reach it. Change it to your real server before clicking anything.
• It's an HTTP client, not a model loader. No VRAM magic here - the model lives on the Ollama server, and this node just forwards your request. If nothing happens, check that Ollama is actually serving before you blame the node.
• SSRF, for real. The pack ships server-side routes that fetch the URL you supply, so anyone who can reach your ComfyUI can point it at internal hosts. If ComfyUI is exposed past your own machine, set the allowlist before starting it:

  export SIMPLE_OLLAMA_ALLOWED_HOSTS="localhost,127.0.0.1,192.168.1.100"

  The connection node honors the same allowlist, and the error message tells you exactly which host to add.

There are bigger, fancier Ollama packs out there - this one is the unglamorous part of a clean setup. Sometimes that's exactly the node you want.